Roofline

Roofline analysis for arithmetic intensity vs achieved performance

Usage

from fasterbench.roofline import RooflineAnalyzer

ra = RooflineAnalyzer(model, sample)
ra.profile(device="cuda")
ra.summary()
fig = ra.plot()
fig.show()

This is a measurement primitive: it measures, it never prescribes. The decisions that consume ra.results live in the compression workflow that calls it.

---

See Also

• Per-layer profiling - Generic per-layer hook infrastructure reused here
• Compute metrics - Model-level FLOPs counting
• Speed metrics - Latency measurement